A Note on Improving Variational Estimation for Multidimensional Item Response Theory

Area of Science:

• Psychometrics
• Statistical modeling
• Social science research

Background:

• Multidimensional item response theory (MIRT) is crucial for analyzing complex constructs in social sciences.
• Estimating MIRT models is computationally intensive, limiting practical application.
• Existing variational estimation methods, like GVEM, offer speed but may introduce bias in discrimination parameters.

Purpose of the Study:

• To address the bias in discrimination parameters observed in variational estimation methods for MIRT.
• To propose and evaluate an enhanced variational estimation algorithm for MIRT models.
• To improve the accuracy and efficiency of parameter estimation in complex MIRT models.

Main Methods:

• Development of an importance-weighted version of the Gaussian variational expectation-maximization (GVEM) algorithm, termed IW-GVEM.
• Integration of adaptive moment estimation for automatic learning rate updates in gradient descent.
• Simulation studies to compare the performance of IW-GVEM against standard GVEM.

Main Results:

• IW-GVEM effectively corrects bias in discrimination parameters compared to standard GVEM.
• The proposed method introduces only a modest increase in computation time.
• The method demonstrates improved accuracy in estimating MIRT models.

Conclusions:

• IW-GVEM offers a more accurate and computationally feasible approach for estimating MIRT models.
• This advancement can facilitate the wider application of MIRT in social science research.
• The IW-GVEM approach may offer insights for improving other psychometric models.